A curated, continuously updated collection of research on Large Language Models for Recommendation — spanning Generative Recommendation and Agentic Recommendation
🔄 Loading latest papers from ArXiv...
A curated collection of seminal papers that have shaped the two main tracks of LLM4Rec: Generative Recommendation and Agentic Recommendation.
Reformulates recommendation as a sequence generation problem — encoding items as (semantic) IDs or tokens and training generative/transformer models to directly decode the next item(s) a user will interact with, replacing the traditional embed-and-retrieve pipeline.
P5 is among the first works to unify diverse recommendation tasks (rating prediction, sequential recommendation, explanation generation, review summarization, direct recommendation) into a single text-to-text generative framework built on a pretrained language model, establishing the "Recommendation as Language Processing" paradigm.
TIGER proposes generative retrieval for recommendation: items are mapped to semantically meaningful Semantic IDs via RQ-VAE, and a Transformer sequence-to-sequence model autoregressively decodes the Semantic ID of the next item, replacing nearest-neighbor search over a unified embedding space.
This Meta paper reformulates industrial-scale recommendation as a sequential transduction task and proposes HSTU, a new generative-recommender architecture that scales to trillions of parameters, outperforms FlashAttention2-based Transformers, and empirically follows a power-law scaling trend with training compute.
OneRec replaces Kuaishou's cascaded retrieve-and-rank pipeline with a single end-to-end generative model: an encoder-decoder with sparse Mixture-of-Experts that performs session-wise generation, further improved with iterative preference alignment, and is deployed in production on a short-video platform with hundreds of millions of daily active users.
Treats recommendation as a reasoning and decision-making process carried out by LLM-powered agents — equipped with memory, tools, planning, and multi-agent collaboration — rather than a single-shot scoring/ranking function.
The evolution of recommender systems has shifted preference storage from rating matrices and dense embeddings to semantic memory in the agentic era...
This work proposes iAgent, an LLM-based agent that acts as a protective shield between users and recommender systems, enabling more transparent and controllable recommendation processes.
MACRec introduces a multi-agent collaboration framework where multiple LLM agents work together to improve recommendation quality through collaborative reasoning and information sharing.
This paper explores the potential of generative agents powered by large language models in recommendation scenarios, discussing their capabilities, challenges, and future directions.
This work presents a framework for simulating realistic user behaviors using LLM-based agents, enabling better testing and evaluation of recommender systems without real user data.
RAH proposes a human-centric framework that leverages LLMs as assistants to facilitate better interaction between users and recommender systems, emphasizing user control and interpretability.
RecMind introduces a comprehensive agent framework powered by large language models that can understand user intent, reason about preferences, and generate personalized recommendations through multi-step reasoning.
InteRecAgent presents an AI agent that integrates large language models to enable interactive and conversational recommendations, supporting natural language queries and explanations for better user engagement.
Pre-extracted multimodal features for MovieLens 1M, enabling research on visual, textual, and acoustic perception in recommendation systems. Derived from movie trailers and metadata, processed with state-of-the-art encoders.
| Modality | Encoder | Dimension | Size |
|---|---|---|---|
| Visual | ResNet50 (ImageNet pre-trained) | 3,706 × 1,000 | ~30 MB |
| Textual | Sentence-Transformers | 3,706 × 384 | ~12 MB |
| Acoustic | VGGish (YouTube-8M pre-trained) | 3,706 × 128 | ~4 MB |
| Interactions | User-Item ratings with metadata | 1,000,209 records | ~39 MB |
Download Components:
Note: All features are stored in NumPy .npy format (float32).
The interaction file follows RecBole format with fields:
userID, itemID, rating, timestamp, x_label (sensitive attribute),
u_gender, u_age, u_occupation.
Total dataset size: ~85 MB.
Citation:
@article{chen2025causality,
title = {Causality-Inspired Fair Representation Learning for Multimodal Recommendation},
author = {Chen, Weixin and Chen, Li and Ni, Yongxin and Zhao, Yuhan},
journal = {ACM Transactions on Information Systems},
year = {2025},
volume = {43},
number = {6},
articleno = {153},
numpages = {29},
doi = {10.1145/3706602}
}
Receive weekly research highlights on LLM-powered recommendation — generative & agentic.
- researchers subscribed | Unsubscribe